What is Agent Visit Optimization (AVO)?

Where does AVO sit next to SEO, GEO, AEO and ACO?

Each discipline answers a different question about the same store. SEO decides whether it is found. GEO and AEO decide whether an AI answer cites it. Agentic commerce optimization (ACO) decides whether an agent recommends it. AVO decides whether the agent that arrives completes the purchase.

The first three are about being chosen. They work on what a search engine or a language model reads before anyone visits: rankings, citations, product feeds and structured data. GEO has a research paper behind the name. ACO is a newer label used in industry guides about feeds and commerce protocols.

AVO starts where they stop. The agent has already picked the store. It is now on the product page, the cart and the checkout. What it finds there decides the sale.

What does an AI agent's visit look like?

The agents observed so far visit in a real browser. They load pages, run the page's scripts, open products, fill forms and reach the cart. And they move differently from people: the pointer jumps instead of traveling, and form fields fill without a single keystroke. One also fetched some pages as plain text.

Don't shopping agents use feeds and protocols instead of pages?

Some may. The ones observed here did not. Protocols such as ACP and UCP, and files such as llms.txt, give an agent a channel beside the pages. In Cromanion's tests, Muse did not use the store's commerce protocol endpoint or its llms.txt. It navigated the storefront like a person.

Advice on agentic commerce often concentrates on the channel beside the pages. The Agentic Commerce Protocol (ACP) was developed by Stripe and OpenAI. The Universal Commerce Protocol (UCP) was co-developed by Google with retailers including Shopify. llms.txt is a proposed file that summarizes a site for language models. Each is real, and some agents use them.

The observation here is narrower, and it is stated as one. Muse was tested once for this, on 2026-09-21: it ignored both the protocol endpoint and llms.txt, and browsed. Grok's shopping agent, in three runs, loaded the storefront pages and went through the checkout form. That is two agents, not the whole market. It is enough to say the visit still happens, and that for these agents the page is the source.

Why does the visit decide the sale?

Because the agent decides on what it can read during the visit, and it has nobody to ask. A person who cannot find the delivery price might guess, call or buy anyway. An agent with a budget and a deadline stops, reports what it found, or leaves for another store.

On 2026-09-24, Muse and Grok's shopping agent were each given the same mission on Cromanion's test store: a pair of headphones with a battery, budget and returns requirement, stopping before payment. The store was built with known problems. These are the ones the agents met, in their own reports:

How is AVO measured?

By the Agent Checkout Rate: the share of AI-agent visits that reach checkout. Payment is often handed back to the person, so checkout is the last step a store can reliably observe. Comparing agents a system acted for with a randomly held-back control group is what separates an effect from a coincidence.

Cromanion counts it in production, on agent visits only, with its permanent holdout as the control group. Raw counts are shown from the first visit. A rate appears only once each group has 30 agent visits, and it is never presented as a lift. The Agent Checkout Rate article explains what counts, and what the number cannot see.

What is AVO not?

It is not a separate store for machines. Every page an agent reads is the page a person reads, with the same products, prices and claims. Serving agents different facts from people is cloaking, and it is outside AVO by definition.

It is also not a claim that agents replace people. In Cromanion's measurement, agents are counted as their own population, apart from people. The point of the term is that this population exists, visits in a browser, and can be measured on its own terms.